A new Markov-chain-related statistical approach for modelling synthetic wind power time series

Abstract

The integration of rising shares of volatile wind power in the generation mix is a major challenge for the future energy system. To address the uncertainties involved in wind power generation, models analysing and simulating the stochastic nature of this energy source are becoming increasingly important. One statistical approach that has been frequently used in the literature is the Markov chain approach. Recently, the method was identified as being of limited use for generating wind time series with time steps shorter than 15-40 min as it is not capable of reproducing the autocorrelation characteristics accurately. This paper presents a new Markov-chain-related statistical approach that is capable of solving this problem by introducing a variable second lag. Furthermore, additional features are presented that allow for the further adjustment of the generated synthetic time series. The influences of the model parameter settings are examined by meaningful parameter variations. The suitability of the approach is demonstrated by an application analysis with the example of the wind feed-in in Germany. It shows that-in contrast to conventional Markov chain approaches-the generated synthetic time series do not systematically underestimate the required storage capacity to balance wind power fluctuation.

Modeling long correlation times using additive binary Markov chains: Applications to wind generation time series

“…As input we use the renewables.ninja [38] wind data set for Germany, which consists of hourly wind capacity factors (i.e., wind power normalized by the rated capacity) based on MERRA reanalysis data [39] simulating the 2014 fleet of wind farms for 1985-2014. A wind generation time series exhibits a seasonal as well as a diurnal periodicity [20,29,30,40,41]: The generation in Germany is usually higher in winter than in summer. The diurnal wind power variation weakly depends on the respective season.…”

“…Periods with below (and above) average wind power generation can last up to weeks [18,19]. Furthermore, the generation can vary significantly on hourly time scales, even when aggregated over large spatial scales [20][21][22][23]. These fluctuations have to be accounted for when designing a future energy system.…”

“…Stochastic models are commonly used to simulate wind speed and wind generation data [20,[26][27][28][29][30]. Typical models include autoregressive moving average and Markov chain * d.witthaut@fz-juelich.de models.…”

“…Simple first-order Markov models often cannot reproduce the temporal characteristics of wind power generation, in particular long correlation times, and therefore they are of very limited use for energy system analyses [26]. Higher-order Markov models require a huge amount of input data and are thus impractical [20,26,27]. Additionally, for ordinary Markov models the data have to be discretized and the model output depends on the exact discretization scheme [20].…”

“…Higher-order Markov models require a huge amount of input data and are thus impractical [20,26,27]. Additionally, for ordinary Markov models the data have to be discretized and the model output depends on the exact discretization scheme [20].…”